Denoising of Low Dose PET Images using a Convolutional Neural Network

LEDESMA, MILAGROS; BELZUNCE, MARTIN A.

San Juan

Congreso; XXIII CONGRESO ARGENTINO DE BIOINGENIERÍA Y XII JORNADAS DE INGENIERÍA CLÍNICA; 2022

SABI

Positron emission tomography (PET) provides quantitative functional images that are key in the study and diagnose of various pathologies, such as cancer and neurodegenerative diseases. However, PET suffers from statistical noise due to the limited number of detected events during the acquisition of this imaging modality. In addition, there is a continues effort to reduce the radiotracer dose injected to the patients, which results in even higher noise levels. In this work, we aimed to investigate the use of a convolutional neural network (CNN) to denoise PET images and compensate for the increase of noise in low dose scans. To achieve this, we implemented a CNN with a Residual UNet architecture that was trained using realistic simulations of PET brain images and evaluated its performance for the simulated data and a real scan. The proposed CNN outperformed two different Gaussian filters, showing promising results that would allow the possibility of performing low dose PET scans in the future.